Vals AI 的退款测试发现 Grok 在15天内发送了187条消息。

该研究于10月18日发表，将六个个人助理代理置于一个模拟的退款争议中，并追踪它们在索赔截止日期过后所做的行为。

By · Published · Updated

Primary source: Vals AI

为何重要

随着助手获得对电子邮件和外部服务的访问权限，产品团队在评估任务完成情况的同时，还需要测试持久性、隐私和升级。Vals 的 one-run-per-model 研究指出了具体风险，但并未说明这些风险发生的频率。

Vals AI's refund test found Grok sent 187 messages in 15 days — Vals AI's six-agent simulation probes what happens when a personal assistant pursues a goal after the legitimate route has closed.

Vals AI, co-founded by Rayan Krishnan (@RayanKrishnan) and Langston Nashold (@langstonnashold), published a study on October 18th reporting that one of six personal-assistant agents sent 187 messages in 15 simulated days while pursuing a 350-pound train refund the user was no longer eligible to claim. The study page carries an October 15th byline date.…

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

Reader comments

Conversation for this story loads after sign-in.